Explainable Machine Learning Model for Predicting Persistent Sepsis-Associated Acute Kidney Injury: Development and Validation Study.

Summary

Using 46,097 sepsis patients across retrospective and prospective cohorts, a 12-variable GBM model predicted persistent SA-AKI with AUCs of 0.87–0.98 across validations and outperformed urinary CCL14 in a prospective cohort. The model is explainable (SHAP) and deployed as a web tool for bedside risk stratification.

Key Findings

• Final GBM model with 12 routine variables achieved AUC 0.870 (internal), 0.891 (MIMIC-III subset), 0.932 (eICU), and 0.983 (single-center external retrospective).
• In a prospective cohort, the GBM (AUC 0.852) outperformed urinary CCL14 (AUC 0.821) for predicting persistent SA-AKI.
• Model explainability via SHAP highlighted AKI stage, ΔCreatinine, urine output, and diuretic dose as top contributors; a web tool was released.

Clinical Implications

Supports early nephrology consultation, conservative nephrotoxin use, and fluid/diuretic stewardship in patients flagged high risk for persistent SA-AKI, beyond reliance on biomarkers alone.

Limitations

• Observational data may harbor residual confounding and site-specific practice biases.
• Generalizability to non-participating healthcare systems and low-resource settings requires further testing.

Future Directions

Prospective impact studies to test whether model-guided care reduces persistent SA-AKI and dialysis; adaptation and validation in low-resource ICUs.